From Semiconductors to Superconductors: Periodic Labs Uses AI to Accelerate Materials Discovery

Materials science has long faced a fundamental bottleneck: the cycle from theoretical prediction to experimental validation often takes years or even decades. In a crossover episode of the Science and Engineering podcast, Liam Fedus and Ekin Dogus Cubuk of Periodic Labs discussed in detail how to use AI to synthesize superintelligence to compress this cycle — extending from semiconductors to superconductors, with the goal aimed directly at a paradigm shift in materials discovery.

Why Materials Discovery Needs AI

Traditional materials R&D relies on density functional theory (DFT) calculations plus experimental trial and error, and there is a huge efficiency gap between the two. DFT can predict the properties of candidate materials, but the computational cost is extremely high; experimental validation is even slower, and a new material often takes years from hypothesis to synthesis.

Cubuk previously participated in the GNoME project at Google DeepMind, which predicted hundreds of thousands of stable crystal structures at once. Fedus came from OpenAI, where he led multiple large-scale model training projects. Their combination represents a new approach in the AI for Science field: not using AI to assist a single step, but using foundation models to reconstruct the entire discovery workflow.

Periodic Labs' core claim is that materials discovery can be modeled as a search problem, and the efficiency of the search depends on the accuracy and coverage of the predictive model. When a model can simultaneously predict structural stability, synthesis feasibility, and target properties, the hit rate of experimental validation will increase substantially.

Density functional theory (DFT) is a cornerstone method in the field of materials computation, predicting a material's electronic structure, mechanical properties, and thermodynamic stability by solving for the electron density distribution. Its advantage is that it does not rely on experimental parameters and can provide first-principles predictions based only on the types and arrangements of atoms. However, the computational complexity of DFT grows cubically or even faster with the number of electrons, and a single calculation for a unit cell containing dozens of atoms may take hours to days. If thousands or tens of thousands of candidate structures need to be screened, the total computational load becomes unbearable. This is precisely where the value of AI models lies: use high-quality data generated by DFT to train surrogate models, and at inference time provide predictions close to DFT accuracy at millisecond speed, thereby expanding the search space from thousands to millions.

GNoME (Graph Networks for Materials Exploration) is a project published by Google DeepMind in Nature in 2023 that uses graph neural networks to predict the stability of inorganic crystal structures at scale. It represents crystals as graphs composed of atoms and chemical bonds, and trains a model to predict formation energy, thereby determining which structures are thermodynamically stable and can be synthesized. GNoME produced about 2.2 million new crystal structures at once, of which about 380,000 were assessed as stable candidates. This order of magnitude far exceeds the total number of inorganic crystals known to humanity over hundreds of years (about tens of thousands), demonstrating the potential of AI for large-scale materials screening. But GNoME also exposed a problem: predicted stability does not equal experimental synthesizability, and a large number of candidate structures lack feasible synthesis pathways. This is exactly the gap that Periodic Labs is trying to fill by introducing synthesis planning.

A Roadmap from Semiconductors to Superconductors

Semiconductors are the first field targeted by Periodic Labs, for a straightforward reason: the computational databases for semiconductor materials are relatively complete, experimental characterization methods are mature, and industrial demand is clear. This allows models to be iteratively validated quickly.

Superconductors are a more ambitious goal. The exploration of room-temperature superconductivity is full of controversy, but even in the field of conventional superconductivity, finding materials with higher critical temperatures and easier processability is still enormously valuable. In the podcast, the two guests mentioned that the complexity of superconducting materials lies in the fact that their properties are highly dependent on electron pairing mechanisms, which places higher demands on the model's physical intuition.

The key to the roadmap lies in transfer learning: can a model trained on semiconductor data generalize to superconductors? Cubuk believes that representation learning of crystal structures is universal, but property prediction requires fine-tuning for different physical mechanisms. This means Periodic Labs needs to build a multi-task framework rather than a single-purpose predictor.

The electron pairing mechanism of superconducting materials is key to understanding their complexity. Conventional superconductors (such as niobium-titanium alloys and MgB₂) follow BCS theory: electrons are indirectly attracted through lattice vibrations (phonons) to form Cooper pairs, thereby conducting current without resistance. The critical temperature Tc is usually relatively low, with the highest at about 39K (MgB₂). For unconventional superconductors (such as copper oxide high-temperature superconductors and iron-based superconductors), there is still no unified theory of the pairing mechanism, which may involve magnetic interactions such as spin fluctuations, making it extremely difficult to directly predict Tc from crystal structure. When dealing with conventional superconductors, AI models can use the existing BCS framework as a prior, but when facing unconventional systems they lack reliable physical equations as constraints and can only rely on data-driven pattern recognition.

The core problem of transfer learning in this scenario is distribution shift. Structures in semiconductor databases are mostly sp³ hybridized systems such as diamond-type and zinc blende-type, while superconductors cover completely different structural families such as perovskites, layered copper oxides, and iron-based compounds. Crystal graph representations (such as atomic distances, coordination numbers, and local symmetry) can be universal across different systems, but the property head needs to be retrained or fine-tuned, because the physical quantities that determine Tc are completely different from those that determine semiconductor band gaps. This means that the shared layers in a multi-task framework are responsible for learning universal crystal representations, while task-specific layers capture their respective physical mechanisms — this architectural design is key to whether Periodic Labs' roadmap can succeed.

The Meaning of Forward Deployed Engineering

The "Forward Deployed Engineering" mentioned in the podcast deserves separate discussion. This concept borrows Palantir's terminology and refers to engineers being directly embedded in customer or experimental teams to quickly adjust models based on real feedback.

In the materials discovery scenario, this means AI researchers need to sit with experimental scientists and understand the actual constraints in the synthesis process — such as temperature windows, precursor compatibility, and equipment limitations. These constraints are difficult to fully obtain from papers, but they are crucial to model usability.

Fedus emphasized that pure computational teams can easily fall into the dilemma of "accurate predictions but impossible synthesis." Forward Deployed Engineering is the organizational means to solve this problem, allowing models to consider experimental reality from the outset.

The Boundaries of Synthesis Superintelligence

The term "Synthesis Superintelligence" can easily cause misunderstanding. The two guests clarified that "superintelligence" here does not mean artificial general intelligence, but rather system capability that surpasses human experts on the specific task of synthesis planning.

Synthesis planning is one of the hardest steps in materials chemistry. Given a target material, how should precursors, reaction pathways, and conditions be chosen? Human chemists rely on experience and intuition, while AI needs to learn from limited reaction data. Cubuk pointed out that current models already perform reasonably well on retrosynthesis tasks, but in the field of inorganic materials they still lag behind organic chemistry.

The boundaries are also reflected in the validation step. AI can generate candidate materials, but the final judgment must come from experiments. Periodic Labs' strategy is to build an automated experimental platform that closes the loop between synthesis and characterization, allowing the model to learn quickly from failures.

Retrosynthesis is a core computational problem in synthesis planning: given a target molecule or material, infer feasible precursors and reaction pathways backward. In the field of organic chemistry, this task already has relatively mature AI solutions, such as template-based methods and sequence-to-sequence model-based approaches. The rules of organic reactions are relatively clear (functional group transformations, protecting group strategies, etc.), and training data are abundant — databases such as Reaxys and SciFinder contain tens of millions of organic reaction records.

Retrosynthesis of inorganic materials is much more difficult. Inorganic reactions often involve solid-state reactions, high-temperature sintering, hydrothermal synthesis, and other pathways. Reaction conditions (temperature, pressure, atmosphere) have a huge impact on the product, and side reactions and intermediate states are difficult to observe. The scale of inorganic reaction databases is far smaller than that of the organic field, and the degree of standardization is also low. More critically, "feasible pathways" in inorganic synthesis are often not unique, and the same material can be obtained through multiple combinations of precursors and conditions, which increases the ambiguity of model learning. Cubuk pointed out that current models still lag significantly in inorganic retrosynthesis, which is also why Periodic Labs needs to build its own automated experimental platform and accumulate proprietary data through closed-loop feedback — public data are insufficient to support learning for this task.

Implications for AI for Science

Periodic Labs' practice reveals a key trend in AI for Science: moving from tooling to platformization. Early work such as AlphaFold solved a single prediction problem, while the next generation of systems needs to integrate prediction, planning, and experimental execution.

This requires a team to simultaneously possess capabilities in machine learning, computational physics, and experimental chemistry, and the organizational difficulty is far higher than that of pure software projects. The combination of Fedus and Cubuk — one from large model training, one from computational materials — is itself a signal: cross-disciplinary collaboration is a basic requirement for this direction.

Another implication is the data strategy. Materials data are scattered across papers, databases, and internal laboratory records, with a low degree of standardization. Periodic Labs must invest substantial effort in data cleaning and structuring; this work is not glamorous but determines success or failure.

The podcast concluded by mentioning that the evaluation standard for AI for Science should shift from "prediction accuracy" to "experimental hit rate" — that is, how many of the materials suggested by the model are ultimately successfully synthesized and validated. This metric is harsher, but closer to real value.

One-line Summary

Periodic Labs uses AI foundation models to reconstruct the materials discovery workflow, compressing R&D cycles from semiconductors to superconductors.

Paragraph Summary

In the podcast, Liam Fedus and Ekin Dogus Cubuk of Periodic Labs explained their approach to accelerating materials discovery with AI: modeling materials discovery as a search problem and using foundation models to simultaneously predict structural stability, synthesis feasibility, and target properties. The roadmap starts with semiconductors and gradually moves to superconductors. The core challenge is that crystal representations can be universal, but property prediction requires fine-tuning for different physical mechanisms. The team emphasizes Forward Deployed Engineering, embedding AI engineers in experimental teams to understand real constraints. So-called "synthesis superintelligence" focuses on the synthesis planning task and still lags behind organic chemistry in inorganic materials. The article points out that AI for Science is moving from tooling to platformization, and the evaluation standard should shift from prediction accuracy to experimental hit rate.